AI Flavor Remover

Problem

AI-generated drafts often sound overly smooth: neat parallelism, uniform sentence length, and mechanical transitions can trigger high AIGC suspicion in tools such as Zhuque, GPTZero, and similar detectors. Simply deleting “AI words” is unreliable because detectors weigh statistical signals like perplexity, burstiness, topic consistency, and word-frequency distribution. This skill reframes the task from removing AI flavor to injecting human signals: concrete names, dates, numbers, colloquial voice, uncertainty, digressions, and fragments.

How It Works

The package centers on scripts/detect.py, scripts/rewrite.py, and references/style-guide.md. detect.py returns a 0-100 human-signal score plus six sub-scores for specificity, burstiness, doubt, digression, voice, and imperfection. rewrite.py performs a conservative surface pass, removing dead filler, mechanical connectors, and double em dashes without deleting substance. The workflow is to score the draft, identify missing human signals, run the light cleanup, rewrite semantically using the style guide, re-score, and then verify on the detector’s official site. For education writing, it preserves required domain terms such as “立德树人” and “核心素养” while rewriting the surrounding formulaic phrasing.

Limits and Notes

The built-in score is not a Zhuque proxy and does not guarantee passing AIGC checks. Detectors keep changing, and official sites remain authoritative. If a result stays high, continue adding specificity, conversational texture, and imperfect syntax rather than relying on synonym swaps alone.